Causal discovery

As explained in Causal AI

Causal discovery is the identification of causal relationships from data, including determining what causes what.

Causal discovery identifies causal relationships from data by learning a causal directed acyclic graph (DAG). Unlike prediction, it asks what causes what, using statistical constraints and explicit assumptions rather than treating associations as interventions. section 1.0 section 4.6

Common discovery strategies include:

• Constraint-based algorithms infer graph structure from conditional independences in the data. section 4.6.1
• Score-based algorithms compare DAGs as explanatory models and seek high goodness-of-fit scores. section 4.6.1
• Continuous-optimization methods relax the discrete DAG search space, while prior knowledge constrains possible edges. section 4.6.1
• Experimental data can improve discovery when intervened variables are identified explicitly. section 4.6.1

The central reasoning pattern is to use causal structure to predict conditional independence. If observed variables violate an independence implied by a candidate DAG, the DAG becomes harder to defend. Faithfulness connects observed independences to d-separations, but Markov-equivalent DAGs can remain indistinguishable, and latent variables can produce serious errors. section 4.1 section 4.7

Blood type and conditional independence

A paternal grandfather’s blood type is an indirect cause of a grandchild’s blood type, yet becomes conditionally independent of it after conditioning on the father’s blood type. This illustrates the kind of constraint discovery algorithms can test. section 4.1

Figure 4.1 Causality induces conditional independence. Your blood type is conditionally independent of your paternal grandfather’s blood type (an indirect cause), given your father’s blood type (a direct cause).

Use the blood-type DAG to visualize how a direct cause blocks information from an indirect ancestor. section 4.1

Treat discovery as exploratory data analysis within a human-led DAG-building workflow, not as an automatic replacement for domain knowledge, model criticism, or downstream causal inference. In Python, PyWhy tools such as causal-learn and DoDiscover provide starting points. section 4.6.4 section 4.6.1

How Causal AI covers it

How deep it goes

Explains algorithm families, conditional-independence foundations, equivalence, and failure modes, but stops short of reliable automatic graph recovery.

Examples

Uses blood-type ancestry to illustrate conditional independence and Netflix churn to frame causal questions.

What it assumes

Assumes causal DAGs and conditional independence while teaching discovery’s role, assumptions, algorithms, and limitations.

What you can do after

Evaluate/understand discovery outputs as imperfect evidence for constructing and stress-testing a causal DAG.

How it connects

Connects causal DAGs with conditional independence, d-separation, faithfulness, latent variables, prior knowledge, and downstream causal inference.

liveBookmanning.com

Read Causal AI on liveBook Get the book at manning.com

Related concepts

• causal graphical modelproduces · A causal graphical model is a probabilistic or structural model whose graph represents causal relationships and whose mechanisms support reasoning about interventions.
• data explorationuses · Data exploration is the systematic examination and summarization of data to understand its structure, quality, distributions, relationships, patterns, and anomalies.
• collideraddressed by · A collider is a variable in a causal graph that receives directed edges from at least two other variables, potentially inducing associations when conditioned on.

This explainer was generated by Manning's knowledge graph from the text of Causal AI. The section links point to the parts of the book it draws on.